Biological reasoning / task collections
Biological reasoning.
Verifiable answers.
Biological reasoning tasks for AI agents, with deterministic scoring.
01 / CellSim · Functional genomics
10 tasksGeneBench-Pro
CRISPRi screens.
Identify a gene knockdown and estimate its effect from a simulated CRISPRi screen.
Can the agent recover the right answer?
Whole-task success: correct gene and effect estimate.
| Problem | Description |
|---|---|
| dADP screen | Identify the strongest dADP-lowering knockdown in a supplied CRISPRi screen and estimate its effect. |
| Acyl-CoA screen | Identify the gene knockdown that lowers a mitochondrial acyl-CoA reporter and estimate its effect. |
| C14:0 PE screen | Identify the gene knockdown that lowers the cytosolic C14:0 PE reporter and estimate its effect. |
| Aminopropionaldehyde screen | Identify the gene knockdown that lowers cytosolic beta-aminopropionaldehyde and estimate its effect. |
| Heptadecanoate screen | Identify the gene knockdown that lowers cytosolic heptadecanoate and estimate its effect. |
| Leukotriene B4 screen | Identify the strongest leukotriene B4-lowering knockdown in a supplied CRISPRi screen and estimate its effect. |
| Protein methyl-lysine screen | Identify the strongest knockdown lowering a cytosolic protein methyl-lysine reporter and estimate its effect. |
| GT1b screen | Identify the strongest GT1b-lowering knockdown in a supplied CRISPRi screen and estimate its effect. |
| L-Homoserine screen | Identify the strongest L-homoserine-lowering knockdown in a supplied CRISPRi screen and estimate its effect. |
| Oleoyl GPE screen | Identify the strongest knockdown lowering the oleoylglycerophosphoethanolamine reporter and estimate its effect. |
02 / Physim · Whole-body physiology
10 tasksSeveral interventions.
One coupled system.
Predict five physiological effects of combined interventions, scored against a simulated matched control.
Can the agent follow the consequences?
Accuracy across the five requested output directions.
| Problem | Description |
|---|---|
| Meat Bro | Predict the combined effects of pressure, hormone changes, arterial stiffness, and liver impairment. |
| Summit Air Scrubber | Predict the combined effects of altitude, exercise, and altered breathing gas. |
| Rescue Basket | Predict responses to blood loss, cold, and head-up tilt. |
| Heart with a Shortcut | Predict exercise and recovery responses with weakened heart function and a vascular shunt. |
| Robot Paramedic | Predict responses to saline and mechanical ventilation after blood loss with impaired kidney function. |
| Shunt Shutdown | Predict the effects of closing a vascular shunt while changing adrenergic signaling and oxygen supply. |
| Oxygen-Loop Fault | Predict hormonal and metabolic responses to mechanical ventilation with elevated oxygen and carbon dioxide. |
| Centrifuge Saline Bag | Predict the combined effects of increased gravity, head-up tilt, supplemental oxygen, and saline. |
| Sealed Ambulance | Predict responses to concentrated saline and altered ventilation after blood loss with impaired kidney function. |
| Cold Transfer | Predict the combined effects of cooling, head-up tilt, and supplemental oxygen. |
Progress in biology requires models that can follow causes through interacting systems and predict their consequences. Our task generators turn that challenge into training and evaluation problems with adjustable difficulty and verifiable answers—from identifying gene effects in experimental data to predicting how multiple interventions change whole-body physiology.
GeneBench-Pro
Our CellSim generator produces multi-stage scientific analysis tasks in GeneBench-Pro’s functional genomics domain. Each problem presents a simulated CRISPRi screen as guide-level sequencing counts. The agent must identify the strongest reporter-lowering gene knockdown and estimate its effect, making analytical decisions that determine whether it reaches the correct conclusion.
The generator draws on CellSim’s metabolic reaction network, spanning 487 reporter metabolites and 1,789 candidate genes. Varying the reporter, gene library, and simulated measurements produces new datasets with controlled difficulty. Each reference answer is checked for recoverability from the supplied data before release. Delivery: hundreds of tasks per week.
Physim
Physim asks what happens when several things change in the same person at once. Interventions affecting blood pressure, hormones, organ function, and metabolism can reinforce or oppose one another. The agent must trace these effects across interacting organ systems and predict their combined consequences.
Each task specifies the starting physiology, interventions, outputs, and observation time. Reference answers come from numerical simulations of the intervention and a matched control. Difficulty varies with the number of interventions and targets, the depth of downstream effects, and the interactions between competing influences. Delivery: 500 tasks in one week; 2,500 tasks within two weeks.